NVIDIA releases open synthetic datasets and tooling to support agentic AI development

NVIDIA published an open data suite under the Nemotron brand to support the development of AI agents, arguing that agentic systems require more than model weights to function robustly in real-world conditions. The company states that agents must handle broken API calls, unfamiliar workflows, and multi-step reasoning, which it frames as a data problem rather than solely a modeling challenge.

The Nemotron open data suite includes over 10 trillion pre-training tokens and millions of post-training samples spanning multiple domains such as general text, code, math, and synthetic data. NVIDIA highlights that synthetic data plays a role in enhancing datasets like Common Crawl and improving reasoning through synthetic math questions, positioning it as a key enabler for scaling agentic data.

To make the post-training data more accessible, NVIDIA released the Nemotron Post-Training v3 Prompt Atlas, an interactive visualization that maps prompt samples from the Nemotron v3 post-training collection. The tool allows users to filter and reorganize samples by dataset, pipeline stage, domain, or tool use, helping developers inspect data mixtures and curate evaluations or fine-tune models.

NVIDIA also introduced Nemotron-Personas, a collection of synthetic personas designed to reflect regional demographic and geographic diversity. Built using NVIDIA’s NeMo Data Designer, the personas aim to help developers test whether their systems serve the intended users, languages, regions, and occupations. As of the announcement, the collection spans ten countries and represents more than 2.4 billion people.

The company argues that synthetic data allows organizations to preserve useful signals without exposing proprietary workflows, customer patterns, or sensitive data, thereby enabling collaboration across companies, researchers, governments, and communities. It frames synthetic data as a tool to build trust between organizations by reducing the need to share sensitive real-world data.

NVIDIA acknowledges tradeoffs in using synthetic data, emphasizing that it should be integrated into a broader system of data sources with clear documentation of generation methods, grounding, lineage, curation, evaluation, and human judgment. It introduces the concept of 'synthetic thresholds' to describe points where synthetic data can no longer be treated as purely real, advocating for transparency in data provenance and evaluation practices.

The announcement coincides with NVIDIA’s broader push to highlight open models and datasets at events like the International Conference on Machine Learning (ICML), where nearly 145 papers cited Nemotron models and datasets. NVIDIA also hosted a livestream titled 'Why Open Data Matters' on July 7, 2026, to discuss the initiative.